regnety_008

Implementation of RegNet proposed in Designing Network Design Spaces

The main idea is to start with a high dimensional search space and iteratively reduce the search space by empirically apply constrains based on the best performing models sampled by the current search space.

The resulting models are light, accurate, and faster than EfficientNets (up to 5x times!)

For example, to go from $AnyNet_A$ to $AnyNet_B$ they fixed the bottleneck ratio $b_i$ for all stage $i$. The following table shows all the restrictions applied from one search space to the next one.

The paper is really well written and very interesting, I highly recommended read it.

ResNet.regnetx_002()
ResNet.regnetx_004()
ResNet.regnetx_006()
ResNet.regnetx_008()
ResNet.regnetx_016()
ResNet.regnetx_040()
ResNet.regnetx_064()
ResNet.regnetx_080()
ResNet.regnetx_120()
ResNet.regnetx_160()
ResNet.regnetx_320()
# Y variants (with SE)
ResNet.regnety_002()
# ...
ResNet.regnetx_320()

You can easily customize your model

Examples:

# change activation
RegNet.regnetx_004(activation = nn.SELU)
# change number of classes (default is 1000 )
RegNet.regnetx_004(n_classes=100)
# pass a different block
RegNet.regnetx_004(block=RegNetYBotteneckBlock)
# change the steam
model = RegNet.regnetx_004(stem=ResNetStemC)
change shortcut
model = RegNet.regnetx_004(block=partial(RegNetYBotteneckBlock, shortcut=ResNetShorcutD))
# store each feature
x = torch.rand((1, 3, 224, 224))
# get features
model = RegNet.regnetx_004()
# first call .features, this will activate the forward hooks and tells the model you'll like to get the features
model.encoder.features
model(torch.randn((1,3,224,224)))
# get the features from the encoder
features = model.encoder.features
print([x.shape for x in features])
#[torch.Size([1, 32, 112, 112]), torch.Size([1, 32, 56, 56]), torch.Size([1, 64, 28, 28]), torch.Size([1, 160, 14, 14])]